File size: 5,883 Bytes
dbac20f |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 |
# Reference: # https://github.com/bytedance/Make-An-Audio-2
import torch
import torch.nn as nn
import torchaudio
from einops import rearrange
from librosa.filters import mel as librosa_mel_fn
def dynamic_range_compression_torch(x, C=1, clip_val=1e-5, norm_fn=torch.log10):
return norm_fn(torch.clamp(x, min=clip_val) * C)
def spectral_normalize_torch(magnitudes, norm_fn):
output = dynamic_range_compression_torch(magnitudes, norm_fn=norm_fn)
return output
class STFTConverter(nn.Module):
def __init__(
self,
*,
sampling_rate: float = 16_000,
n_fft: int = 1024,
num_mels: int = 128,
hop_size: int = 256,
win_size: int = 1024,
fmin: float = 0,
fmax: float = 8_000,
norm_fn=torch.log,
):
super().__init__()
self.sampling_rate = sampling_rate
self.n_fft = n_fft
self.num_mels = num_mels
self.hop_size = hop_size
self.win_size = win_size
self.fmin = fmin
self.fmax = fmax
self.norm_fn = norm_fn
mel = librosa_mel_fn(sr=self.sampling_rate,
n_fft=self.n_fft,
n_mels=self.num_mels,
fmin=self.fmin,
fmax=self.fmax)
mel_basis = torch.from_numpy(mel).float()
hann_window = torch.hann_window(self.win_size)
self.register_buffer('mel_basis', mel_basis)
self.register_buffer('hann_window', hann_window)
@property
def device(self):
return self.hann_window.device
def forward(self, waveform: torch.Tensor) -> torch.Tensor:
# input: batch_size * length
bs = waveform.shape[0]
waveform = waveform.clamp(min=-1., max=1.)
spec = torch.stft(waveform,
self.n_fft,
hop_length=self.hop_size,
win_length=self.win_size,
window=self.hann_window,
center=True,
pad_mode='reflect',
normalized=False,
onesided=True,
return_complex=True)
spec = torch.view_as_real(spec)
# print('After stft', spec.shape, spec.min(), spec.max(), spec.mean())
power = spec.pow(2).sum(-1)
angle = torch.atan2(spec[..., 1], spec[..., 0])
print('power', power.shape, power.min(), power.max(), power.mean())
print('angle', angle.shape, angle.min(), angle.max(), angle.mean())
# print('mel', self.mel_basis.shape, self.mel_basis.min(), self.mel_basis.max(),
# self.mel_basis.mean())
# spec = rearrange(spec, 'b f t c -> (b c) f t')
# spec = self.mel_transform(spec)
# spec = torch.matmul(self.mel_basis, spec)
# print('After mel', spec.shape, spec.min(), spec.max(), spec.mean())
# spec = spectral_normalize_torch(spec, self.norm_fn)
# print('After norm', spec.shape, spec.min(), spec.max(), spec.mean())
# compute magnitude
# magnitude = torch.sqrt((spec**2).sum(-1))
# normalize by magnitude
# scaled_magnitude = torch.log10(magnitude.clamp(min=1e-5)) * 10
# spec = spec / magnitude.unsqueeze(-1) * scaled_magnitude.unsqueeze(-1)
# power = torch.log10(power.clamp(min=1e-5)) * 10
power = torch.log10(power.clamp(min=1e-5))
print('After scaling', power.shape, power.min(), power.max(), power.mean())
spec = torch.stack([power, angle], dim=-1)
# spec = rearrange(spec, '(b c) f t -> b c f t', b=bs)
spec = rearrange(spec, 'b f t c -> b c f t', b=bs)
# spec[:, :, 400:] = 0
return spec
def invert(self, spec: torch.Tensor, length: int) -> torch.Tensor:
bs = spec.shape[0]
# spec = rearrange(spec, 'b c f t -> (b c) f t')
# print(spec.shape, self.mel_basis.shape)
# spec = torch.linalg.lstsq(self.mel_basis.unsqueeze(0), spec).solution
# spec = torch.linalg.pinv(self.mel_basis.unsqueeze(0)) @ spec
# spec = self.invmel_transform(spec)
spec = rearrange(spec, 'b c f t -> b f t c', b=bs).contiguous()
# spec[..., 0] = 10**(spec[..., 0] / 10)
power = spec[..., 0]
power = 10**power
# print('After unscaling', spec[..., 0].shape, spec[..., 0].min(), spec[..., 0].max(),
# spec[..., 0].mean())
unit_vector = torch.stack([
torch.cos(spec[..., 1]),
torch.sin(spec[..., 1]),
], dim=-1)
spec = torch.sqrt(power) * unit_vector
# spec = rearrange(spec, '(b c) f t -> b f t c', b=bs).contiguous()
spec = torch.view_as_complex(spec)
waveform = torch.istft(
spec,
self.n_fft,
length=length,
hop_length=self.hop_size,
win_length=self.win_size,
window=self.hann_window,
center=True,
normalized=False,
onesided=True,
return_complex=False,
)
return waveform
if __name__ == '__main__':
converter = STFTConverter(sampling_rate=16000)
signal = torchaudio.load('./output/ZZ6GRocWW38_000090.wav')[0]
# resample signal at 44100 Hz
# signal = torchaudio.transforms.Resample(16_000, 44_100)(signal)
L = signal.shape[1]
print('Input signal', signal.shape)
spec = converter(signal)
print('Final spec', spec.shape)
signal_recon = converter.invert(spec, length=L)
print('Output signal', signal_recon.shape, signal_recon.min(), signal_recon.max(),
signal_recon.mean())
print('MSE', torch.nn.functional.mse_loss(signal, signal_recon))
torchaudio.save('./output/ZZ6GRocWW38_000090_recon.wav', signal_recon, 16000)
|